Square Integer Matrix with a Single Non-Integer Entry in Its Inverse

Matrix inversion is one of the most significant operations on a matrix. For any non-singular matrix A∈Zn×n, the inverse of this matrix may contain countless numbers of non-integer entries. These entries could be endless floating-point numbers. Storing, transmitting, or operating such an inverse could be cumbersome, especially when the size n is large. The only square integer matrix that is guaranteed to have an integer matrix as its inverse is a unimodular matrix U∈Zn×n. With the property that det(U)=±1, then U−1∈Zn×n is guaranteed such that UU−1=I, where I∈Zn×n is an identity matrix. In this paper, we propose a new integer matrix G˜∈Zn×n, which is referred to as an almost-unimodular matrix. With det(G˜)≠±1, the inverse of this matrix, G˜−1∈Rn×n, is proven to consist of only a single non-integer entry. The almost-unimodular matrix could be useful in various areas, such as lattice-based cryptography, computer graphics, lattice-based computational problems, or any area where the inversion of a large integer matrix is necessary, especially when the determinant of the matrix is required not to equal ±1. Therefore, the almost-unimodular matrix could be an alternative to the unimodular matrix.

Perron numbers that satisfy Fermat’s equation

In this note, it is shown that if \ell and m are positive integers such that \ell > m, then there is a Perron number \rho such that \rho^n + (\rho + m)^n = (\rho + \ell)^n. It is also shown that there is an aperiodic integer matrix C such that C^n + (C+ m I_n)^n = (C + \ell I_n)^n.

A Note on Decomposable and Reducible Integer Matrices

We propose necessary and sufficient conditions for an integer matrix to be decomposable in terms of its Hermite normal form. Specifically, to each integer matrix, we associate a symmetric integer matrix whose reducibility can be efficiently determined by elementary linear algebra techniques, and which completely determines the decomposability of the first one.

Color Image Encryption Based on Cross 2d Hyperchaotic Map Using Combined Cycle Shift Scrambling and Selecting Diffusion

A novel color image encryption algorithm based on a cross 2D hyperchaotic map is proposed in this paper. The cross 2D hyperchaotic map is constructed using one nonlinear function and two chaotic maps with cross structure. Chaotic behaviors are illustrated using bifurcation diagrams, Lyapunov exponent spectra and phase portraits, etc. In the color image encryption algorithm, the keys are generated using hash function SHA-512 and the information of plain color image. First, the color plain image is converted to a combined bit-level matrix and permuted by the chaos based row and column combined cycle shift scrambling method. Then the scrambled integer matrix is diffused according to the selecting sequence which depends on the chaotic sequence. Last, the cipher color image is obtained by decomposed the diffused matrix. Simulation results show that the algorithm can encrypt the color image effectively and has good security.

On generic complexity of the subset sum problem in monoids and groups of integer matrix of order two

Generic-case approach to algorithmic problems was suggested by A. Miasnikov, I. Kapovich, P. Schupp and V. Shpilrain in 2003. This approach studies behavior of an algo-rithm on typical (almost all) inputs and ignores the rest of inputs. In this paper, we prove that the subset sum problems for the monoid of integer positive unimodular matrices of the second order, the special linear group of the second order, and the modular group are generically solvable in polynomial time.

Overflow Aware Quantization: Accelerating Neural Network Inference by Low-bit Multiply-Accumulate Operations

The inherent heavy computation of deep neural networks prevents their widespread applications. A widely used method for accelerating model inference is quantization, by replacing the input operands of a network using fixed-point values. Then the majority of computation costs focus on the integer matrix multiplication accumulation. In fact, high-bit accumulator leads to partially wasted computation and low-bit one typically suffers from numerical overflow. To address this problem, we propose an overflow aware quantization method by designing trainable adaptive fixed-point representation, to optimize the number of bits for each input tensor while prohibiting numeric overflow during the computation. With the proposed method, we are able to fully utilize the computing power to minimize the quantization loss and obtain optimized inference performance. To verify the effectiveness of our method, we conduct image classification, object detection, and semantic segmentation tasks on ImageNet, Pascal VOC, and COCO datasets, respectively. Experimental results demonstrate that the proposed method can achieve comparable performance with state-of-the-art quantization methods while accelerating the inference process by about 2 times.